LLVM API Documentation
00001 //===---- ScheduleDAGInstrs.cpp - MachineInstr Rescheduling ---------------===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This implements the ScheduleDAGInstrs class, which implements re-scheduling 00011 // of MachineInstrs. 00012 // 00013 //===----------------------------------------------------------------------===// 00014 00015 #define DEBUG_TYPE "sched-instrs" 00016 #include "ScheduleDAGInstrs.h" 00017 #include "llvm/Operator.h" 00018 #include "llvm/Analysis/AliasAnalysis.h" 00019 #include "llvm/CodeGen/MachineFunctionPass.h" 00020 #include "llvm/CodeGen/MachineRegisterInfo.h" 00021 #include "llvm/CodeGen/PseudoSourceValue.h" 00022 #include "llvm/Target/TargetMachine.h" 00023 #include "llvm/Target/TargetInstrInfo.h" 00024 #include "llvm/Target/TargetRegisterInfo.h" 00025 #include "llvm/Target/TargetSubtarget.h" 00026 #include "llvm/Support/Debug.h" 00027 #include "llvm/Support/raw_ostream.h" 00028 #include "llvm/ADT/SmallSet.h" 00029 using namespace llvm; 00030 00031 ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf, 00032 const MachineLoopInfo &mli, 00033 const MachineDominatorTree &mdt) 00034 : ScheduleDAG(mf), MLI(mli), MDT(mdt), LoopRegs(MLI, MDT) {} 00035 00036 /// Run - perform scheduling. 00037 /// 00038 void ScheduleDAGInstrs::Run(MachineBasicBlock *bb, 00039 MachineBasicBlock::iterator begin, 00040 MachineBasicBlock::iterator end, 00041 unsigned endcount) { 00042 BB = bb; 00043 Begin = begin; 00044 InsertPosIndex = endcount; 00045 00046 ScheduleDAG::Run(bb, end); 00047 } 00048 00049 /// getUnderlyingObjectFromInt - This is the function that does the work of 00050 /// looking through basic ptrtoint+arithmetic+inttoptr sequences. 00051 static const Value *getUnderlyingObjectFromInt(const Value *V) { 00052 do { 00053 if (const Operator *U = dyn_cast<Operator>(V)) { 00054 // If we find a ptrtoint, we can transfer control back to the 00055 // regular getUnderlyingObjectFromInt. 00056 if (U->getOpcode() == Instruction::PtrToInt) 00057 return U->getOperand(0); 00058 // If we find an add of a constant or a multiplied value, it's 00059 // likely that the other operand will lead us to the base 00060 // object. We don't have to worry about the case where the 00061 // object address is somehow being computed by the multiply, 00062 // because our callers only care when the result is an 00063 // identifibale object. 00064 if (U->getOpcode() != Instruction::Add || 00065 (!isa<ConstantInt>(U->getOperand(1)) && 00066 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul)) 00067 return V; 00068 V = U->getOperand(0); 00069 } else { 00070 return V; 00071 } 00072 assert(isa<IntegerType>(V->getType()) && "Unexpected operand type!"); 00073 } while (1); 00074 } 00075 00076 /// getUnderlyingObject - This is a wrapper around Value::getUnderlyingObject 00077 /// and adds support for basic ptrtoint+arithmetic+inttoptr sequences. 00078 static const Value *getUnderlyingObject(const Value *V) { 00079 // First just call Value::getUnderlyingObject to let it do what it does. 00080 do { 00081 V = V->getUnderlyingObject(); 00082 // If it found an inttoptr, use special code to continue climing. 00083 if (Operator::getOpcode(V) != Instruction::IntToPtr) 00084 break; 00085 const Value *O = getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0)); 00086 // If that succeeded in finding a pointer, continue the search. 00087 if (!isa<PointerType>(O->getType())) 00088 break; 00089 V = O; 00090 } while (1); 00091 return V; 00092 } 00093 00094 /// getUnderlyingObjectForInstr - If this machine instr has memory reference 00095 /// information and it can be tracked to a normal reference to a known 00096 /// object, return the Value for that object. Otherwise return null. 00097 static const Value *getUnderlyingObjectForInstr(const MachineInstr *MI) { 00098 if (!MI->hasOneMemOperand() || 00099 !MI->memoperands_begin()->getValue() || 00100 MI->memoperands_begin()->isVolatile()) 00101 return 0; 00102 00103 const Value *V = MI->memoperands_begin()->getValue(); 00104 if (!V) 00105 return 0; 00106 00107 V = getUnderlyingObject(V); 00108 if (!isa<PseudoSourceValue>(V) && !isIdentifiedObject(V)) 00109 return 0; 00110 00111 return V; 00112 } 00113 00114 void ScheduleDAGInstrs::StartBlock(MachineBasicBlock *BB) { 00115 if (MachineLoop *ML = MLI.getLoopFor(BB)) 00116 if (BB == ML->getLoopLatch()) { 00117 MachineBasicBlock *Header = ML->getHeader(); 00118 for (MachineBasicBlock::livein_iterator I = Header->livein_begin(), 00119 E = Header->livein_end(); I != E; ++I) 00120 LoopLiveInRegs.insert(*I); 00121 LoopRegs.VisitLoop(ML); 00122 } 00123 } 00124 00125 void ScheduleDAGInstrs::BuildSchedGraph() { 00126 // We'll be allocating one SUnit for each instruction, plus one for 00127 // the region exit node. 00128 SUnits.reserve(BB->size()); 00129 00130 // We build scheduling units by walking a block's instruction list from bottom 00131 // to top. 00132 00133 // Remember where a generic side-effecting instruction is as we procede. If 00134 // ChainMMO is null, this is assumed to have arbitrary side-effects. If 00135 // ChainMMO is non-null, then Chain makes only a single memory reference. 00136 SUnit *Chain = 0; 00137 MachineMemOperand *ChainMMO = 0; 00138 00139 // Memory references to specific known memory locations are tracked so that 00140 // they can be given more precise dependencies. 00141 std::map<const Value *, SUnit *> MemDefs; 00142 std::map<const Value *, std::vector<SUnit *> > MemUses; 00143 00144 // Check to see if the scheduler cares about latencies. 00145 bool UnitLatencies = ForceUnitLatencies(); 00146 00147 // Ask the target if address-backscheduling is desirable, and if so how much. 00148 const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>(); 00149 unsigned SpecialAddressLatency = ST.getSpecialAddressLatency(); 00150 00151 // Walk the list of instructions, from bottom moving up. 00152 for (MachineBasicBlock::iterator MII = InsertPos, MIE = Begin; 00153 MII != MIE; --MII) { 00154 MachineInstr *MI = prior(MII); 00155 const TargetInstrDesc &TID = MI->getDesc(); 00156 assert(!TID.isTerminator() && !MI->isLabel() && 00157 "Cannot schedule terminators or labels!"); 00158 // Create the SUnit for this MI. 00159 SUnit *SU = NewSUnit(MI); 00160 00161 // Assign the Latency field of SU using target-provided information. 00162 if (UnitLatencies) 00163 SU->Latency = 1; 00164 else 00165 ComputeLatency(SU); 00166 00167 // Add register-based dependencies (data, anti, and output). 00168 for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) { 00169 const MachineOperand &MO = MI->getOperand(j); 00170 if (!MO.isReg()) continue; 00171 unsigned Reg = MO.getReg(); 00172 if (Reg == 0) continue; 00173 00174 assert(TRI->isPhysicalRegister(Reg) && "Virtual register encountered!"); 00175 std::vector<SUnit *> &UseList = Uses[Reg]; 00176 std::vector<SUnit *> &DefList = Defs[Reg]; 00177 // Optionally add output and anti dependencies. For anti 00178 // dependencies we use a latency of 0 because for a multi-issue 00179 // target we want to allow the defining instruction to issue 00180 // in the same cycle as the using instruction. 00181 // TODO: Using a latency of 1 here for output dependencies assumes 00182 // there's no cost for reusing registers. 00183 SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output; 00184 unsigned AOLatency = (Kind == SDep::Anti) ? 0 : 1; 00185 for (unsigned i = 0, e = DefList.size(); i != e; ++i) { 00186 SUnit *DefSU = DefList[i]; 00187 if (DefSU != SU && 00188 (Kind != SDep::Output || !MO.isDead() || 00189 !DefSU->getInstr()->registerDefIsDead(Reg))) 00190 DefSU->addPred(SDep(SU, Kind, AOLatency, /*Reg=*/Reg)); 00191 } 00192 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) { 00193 std::vector<SUnit *> &DefList = Defs[*Alias]; 00194 for (unsigned i = 0, e = DefList.size(); i != e; ++i) { 00195 SUnit *DefSU = DefList[i]; 00196 if (DefSU != SU && 00197 (Kind != SDep::Output || !MO.isDead() || 00198 !DefSU->getInstr()->registerDefIsDead(Reg))) 00199 DefSU->addPred(SDep(SU, Kind, AOLatency, /*Reg=*/ *Alias)); 00200 } 00201 } 00202 00203 if (MO.isDef()) { 00204 // Add any data dependencies. 00205 unsigned DataLatency = SU->Latency; 00206 for (unsigned i = 0, e = UseList.size(); i != e; ++i) { 00207 SUnit *UseSU = UseList[i]; 00208 if (UseSU != SU) { 00209 unsigned LDataLatency = DataLatency; 00210 // Optionally add in a special extra latency for nodes that 00211 // feed addresses. 00212 // TODO: Do this for register aliases too. 00213 // TODO: Perhaps we should get rid of 00214 // SpecialAddressLatency and just move this into 00215 // adjustSchedDependency for the targets that care about 00216 // it. 00217 if (SpecialAddressLatency != 0 && !UnitLatencies) { 00218 MachineInstr *UseMI = UseSU->getInstr(); 00219 const TargetInstrDesc &UseTID = UseMI->getDesc(); 00220 int RegUseIndex = UseMI->findRegisterUseOperandIdx(Reg); 00221 assert(RegUseIndex >= 0 && "UseMI doesn's use register!"); 00222 if ((UseTID.mayLoad() || UseTID.mayStore()) && 00223 (unsigned)RegUseIndex < UseTID.getNumOperands() && 00224 UseTID.OpInfo[RegUseIndex].isLookupPtrRegClass()) 00225 LDataLatency += SpecialAddressLatency; 00226 } 00227 // Adjust the dependence latency using operand def/use 00228 // information (if any), and then allow the target to 00229 // perform its own adjustments. 00230 const SDep& dep = SDep(SU, SDep::Data, LDataLatency, Reg); 00231 if (!UnitLatencies) { 00232 ComputeOperandLatency(SU, UseSU, (SDep &)dep); 00233 ST.adjustSchedDependency(SU, UseSU, (SDep &)dep); 00234 } 00235 UseSU->addPred(dep); 00236 } 00237 } 00238 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) { 00239 std::vector<SUnit *> &UseList = Uses[*Alias]; 00240 for (unsigned i = 0, e = UseList.size(); i != e; ++i) { 00241 SUnit *UseSU = UseList[i]; 00242 if (UseSU != SU) { 00243 const SDep& dep = SDep(SU, SDep::Data, DataLatency, *Alias); 00244 if (!UnitLatencies) { 00245 ComputeOperandLatency(SU, UseSU, (SDep &)dep); 00246 ST.adjustSchedDependency(SU, UseSU, (SDep &)dep); 00247 } 00248 UseSU->addPred(dep); 00249 } 00250 } 00251 } 00252 00253 // If a def is going to wrap back around to the top of the loop, 00254 // backschedule it. 00255 if (!UnitLatencies && DefList.empty()) { 00256 LoopDependencies::LoopDeps::iterator I = LoopRegs.Deps.find(Reg); 00257 if (I != LoopRegs.Deps.end()) { 00258 const MachineOperand *UseMO = I->second.first; 00259 unsigned Count = I->second.second; 00260 const MachineInstr *UseMI = UseMO->getParent(); 00261 unsigned UseMOIdx = UseMO - &UseMI->getOperand(0); 00262 const TargetInstrDesc &UseTID = UseMI->getDesc(); 00263 // TODO: If we knew the total depth of the region here, we could 00264 // handle the case where the whole loop is inside the region but 00265 // is large enough that the isScheduleHigh trick isn't needed. 00266 if (UseMOIdx < UseTID.getNumOperands()) { 00267 // Currently, we only support scheduling regions consisting of 00268 // single basic blocks. Check to see if the instruction is in 00269 // the same region by checking to see if it has the same parent. 00270 if (UseMI->getParent() != MI->getParent()) { 00271 unsigned Latency = SU->Latency; 00272 if (UseTID.OpInfo[UseMOIdx].isLookupPtrRegClass()) 00273 Latency += SpecialAddressLatency; 00274 // This is a wild guess as to the portion of the latency which 00275 // will be overlapped by work done outside the current 00276 // scheduling region. 00277 Latency -= std::min(Latency, Count); 00278 // Add the artifical edge. 00279 ExitSU.addPred(SDep(SU, SDep::Order, Latency, 00280 /*Reg=*/0, /*isNormalMemory=*/false, 00281 /*isMustAlias=*/false, 00282 /*isArtificial=*/true)); 00283 } else if (SpecialAddressLatency > 0 && 00284 UseTID.OpInfo[UseMOIdx].isLookupPtrRegClass()) { 00285 // The entire loop body is within the current scheduling region 00286 // and the latency of this operation is assumed to be greater 00287 // than the latency of the loop. 00288 // TODO: Recursively mark data-edge predecessors as 00289 // isScheduleHigh too. 00290 SU->isScheduleHigh = true; 00291 } 00292 } 00293 LoopRegs.Deps.erase(I); 00294 } 00295 } 00296 00297 UseList.clear(); 00298 if (!MO.isDead()) 00299 DefList.clear(); 00300 DefList.push_back(SU); 00301 } else { 00302 UseList.push_back(SU); 00303 } 00304 } 00305 00306 // Add chain dependencies. 00307 // Note that isStoreToStackSlot and isLoadFromStackSLot are not usable 00308 // after stack slots are lowered to actual addresses. 00309 // TODO: Use an AliasAnalysis and do real alias-analysis queries, and 00310 // produce more precise dependence information. 00311 if (TID.isCall() || TID.hasUnmodeledSideEffects()) { 00312 new_chain: 00313 // This is the conservative case. Add dependencies on all memory 00314 // references. 00315 if (Chain) 00316 Chain->addPred(SDep(SU, SDep::Order, SU->Latency)); 00317 Chain = SU; 00318 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k) 00319 PendingLoads[k]->addPred(SDep(SU, SDep::Order, SU->Latency)); 00320 PendingLoads.clear(); 00321 for (std::map<const Value *, SUnit *>::iterator I = MemDefs.begin(), 00322 E = MemDefs.end(); I != E; ++I) { 00323 I->second->addPred(SDep(SU, SDep::Order, SU->Latency)); 00324 I->second = SU; 00325 } 00326 for (std::map<const Value *, std::vector<SUnit *> >::iterator I = 00327 MemUses.begin(), E = MemUses.end(); I != E; ++I) { 00328 for (unsigned i = 0, e = I->second.size(); i != e; ++i) 00329 I->second[i]->addPred(SDep(SU, SDep::Order, SU->Latency)); 00330 I->second.clear(); 00331 } 00332 // See if it is known to just have a single memory reference. 00333 MachineInstr *ChainMI = Chain->getInstr(); 00334 const TargetInstrDesc &ChainTID = ChainMI->getDesc(); 00335 if (!ChainTID.isCall() && 00336 !ChainTID.hasUnmodeledSideEffects() && 00337 ChainMI->hasOneMemOperand() && 00338 !ChainMI->memoperands_begin()->isVolatile() && 00339 ChainMI->memoperands_begin()->getValue()) 00340 // We know that the Chain accesses one specific memory location. 00341 ChainMMO = &*ChainMI->memoperands_begin(); 00342 else 00343 // Unknown memory accesses. Assume the worst. 00344 ChainMMO = 0; 00345 } else if (TID.mayStore()) { 00346 if (const Value *V = getUnderlyingObjectForInstr(MI)) { 00347 // A store to a specific PseudoSourceValue. Add precise dependencies. 00348 // Handle the def in MemDefs, if there is one. 00349 std::map<const Value *, SUnit *>::iterator I = MemDefs.find(V); 00350 if (I != MemDefs.end()) { 00351 I->second->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0, 00352 /*isNormalMemory=*/true)); 00353 I->second = SU; 00354 } else { 00355 MemDefs[V] = SU; 00356 } 00357 // Handle the uses in MemUses, if there are any. 00358 std::map<const Value *, std::vector<SUnit *> >::iterator J = 00359 MemUses.find(V); 00360 if (J != MemUses.end()) { 00361 for (unsigned i = 0, e = J->second.size(); i != e; ++i) 00362 J->second[i]->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0, 00363 /*isNormalMemory=*/true)); 00364 J->second.clear(); 00365 } 00366 // Add dependencies from all the PendingLoads, since without 00367 // memoperands we must assume they alias anything. 00368 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k) 00369 PendingLoads[k]->addPred(SDep(SU, SDep::Order, SU->Latency)); 00370 // Add a general dependence too, if needed. 00371 if (Chain) 00372 Chain->addPred(SDep(SU, SDep::Order, SU->Latency)); 00373 } else 00374 // Treat all other stores conservatively. 00375 goto new_chain; 00376 } else if (TID.mayLoad()) { 00377 if (TII->isInvariantLoad(MI)) { 00378 // Invariant load, no chain dependencies needed! 00379 } else if (const Value *V = getUnderlyingObjectForInstr(MI)) { 00380 // A load from a specific PseudoSourceValue. Add precise dependencies. 00381 std::map<const Value *, SUnit *>::iterator I = MemDefs.find(V); 00382 if (I != MemDefs.end()) 00383 I->second->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0, 00384 /*isNormalMemory=*/true)); 00385 MemUses[V].push_back(SU); 00386 00387 // Add a general dependence too, if needed. 00388 if (Chain && (!ChainMMO || 00389 (ChainMMO->isStore() || ChainMMO->isVolatile()))) 00390 Chain->addPred(SDep(SU, SDep::Order, SU->Latency)); 00391 } else if (MI->hasVolatileMemoryRef()) { 00392 // Treat volatile loads conservatively. Note that this includes 00393 // cases where memoperand information is unavailable. 00394 goto new_chain; 00395 } else { 00396 // A normal load. Depend on the general chain, as well as on 00397 // all stores. In the absense of MachineMemOperand information, 00398 // we can't even assume that the load doesn't alias well-behaved 00399 // memory locations. 00400 if (Chain) 00401 Chain->addPred(SDep(SU, SDep::Order, SU->Latency)); 00402 for (std::map<const Value *, SUnit *>::iterator I = MemDefs.begin(), 00403 E = MemDefs.end(); I != E; ++I) 00404 I->second->addPred(SDep(SU, SDep::Order, SU->Latency)); 00405 PendingLoads.push_back(SU); 00406 } 00407 } 00408 } 00409 00410 for (int i = 0, e = TRI->getNumRegs(); i != e; ++i) { 00411 Defs[i].clear(); 00412 Uses[i].clear(); 00413 } 00414 PendingLoads.clear(); 00415 } 00416 00417 void ScheduleDAGInstrs::FinishBlock() { 00418 // Nothing to do. 00419 } 00420 00421 void ScheduleDAGInstrs::ComputeLatency(SUnit *SU) { 00422 const InstrItineraryData &InstrItins = TM.getInstrItineraryData(); 00423 00424 // Compute the latency for the node. 00425 SU->Latency = 00426 InstrItins.getStageLatency(SU->getInstr()->getDesc().getSchedClass()); 00427 00428 // Simplistic target-independent heuristic: assume that loads take 00429 // extra time. 00430 if (InstrItins.isEmpty()) 00431 if (SU->getInstr()->getDesc().mayLoad()) 00432 SU->Latency += 2; 00433 } 00434 00435 void ScheduleDAGInstrs::ComputeOperandLatency(SUnit *Def, SUnit *Use, 00436 SDep& dep) const { 00437 const InstrItineraryData &InstrItins = TM.getInstrItineraryData(); 00438 if (InstrItins.isEmpty()) 00439 return; 00440 00441 // For a data dependency with a known register... 00442 if ((dep.getKind() != SDep::Data) || (dep.getReg() == 0)) 00443 return; 00444 00445 const unsigned Reg = dep.getReg(); 00446 00447 // ... find the definition of the register in the defining 00448 // instruction 00449 MachineInstr *DefMI = Def->getInstr(); 00450 int DefIdx = DefMI->findRegisterDefOperandIdx(Reg); 00451 if (DefIdx != -1) { 00452 int DefCycle = InstrItins.getOperandCycle(DefMI->getDesc().getSchedClass(), DefIdx); 00453 if (DefCycle >= 0) { 00454 MachineInstr *UseMI = Use->getInstr(); 00455 const unsigned UseClass = UseMI->getDesc().getSchedClass(); 00456 00457 // For all uses of the register, calculate the maxmimum latency 00458 int Latency = -1; 00459 for (unsigned i = 0, e = UseMI->getNumOperands(); i != e; ++i) { 00460 const MachineOperand &MO = UseMI->getOperand(i); 00461 if (!MO.isReg() || !MO.isUse()) 00462 continue; 00463 unsigned MOReg = MO.getReg(); 00464 if (MOReg != Reg) 00465 continue; 00466 00467 int UseCycle = InstrItins.getOperandCycle(UseClass, i); 00468 if (UseCycle >= 0) 00469 Latency = std::max(Latency, DefCycle - UseCycle + 1); 00470 } 00471 00472 // If we found a latency, then replace the existing dependence latency. 00473 if (Latency >= 0) 00474 dep.setLatency(Latency); 00475 } 00476 } 00477 } 00478 00479 void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const { 00480 SU->getInstr()->dump(); 00481 } 00482 00483 std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const { 00484 std::string s; 00485 raw_string_ostream oss(s); 00486 if (SU == &EntrySU) 00487 oss << "<entry>"; 00488 else if (SU == &ExitSU) 00489 oss << "<exit>"; 00490 else 00491 SU->getInstr()->print(oss); 00492 return oss.str(); 00493 } 00494 00495 // EmitSchedule - Emit the machine code in scheduled order. 00496 MachineBasicBlock *ScheduleDAGInstrs::EmitSchedule() { 00497 // For MachineInstr-based scheduling, we're rescheduling the instructions in 00498 // the block, so start by removing them from the block. 00499 while (Begin != InsertPos) { 00500 MachineBasicBlock::iterator I = Begin; 00501 ++Begin; 00502 BB->remove(I); 00503 } 00504 00505 // Then re-insert them according to the given schedule. 00506 for (unsigned i = 0, e = Sequence.size(); i != e; i++) { 00507 SUnit *SU = Sequence[i]; 00508 if (!SU) { 00509 // Null SUnit* is a noop. 00510 EmitNoop(); 00511 continue; 00512 } 00513 00514 BB->insert(InsertPos, SU->getInstr()); 00515 } 00516 00517 // Update the Begin iterator, as the first instruction in the block 00518 // may have been scheduled later. 00519 if (!Sequence.empty()) 00520 Begin = Sequence[0]->getInstr(); 00521 00522 return BB; 00523 }